For years, if you wanted access to OpenAI's most capable artificial intelligence, you had one option: pay for it through their cloud. That changed in August 2025, when the company released GPT-OSS — a family of open-weight models that anyone can download, run locally, and use for free.
It's a significant shift, and one worth understanding even if you've never written a line of code in your life.
What exactly has OpenAI released?
GPT-OSS comes in two sizes. The larger model, GPT-OSS-120B, has 117 billion parameters and performs close to OpenAI's own paid o4-mini model on reasoning tasks. The smaller version, GPT-OSS-20B, delivers results comparable to o3-mini — and can run on a laptop with just 16 gigabytes of memory.
Both models use a clever architecture called "mixture of experts," which means only a fraction of the model activates for any given task. That's what makes them so efficient: the 20B model activates just 3.6 billion parameters per query, keeping it fast and light enough for consumer hardware.
"Nothing of this quality has come close to that speed," said Dustin Carr, co-founder and CTO of AI startup Darkviolet.ai, who reported the 20B model running at 45 to 50 tokens per second on his MacBook.
Why the licence matters
The models are released under the Apache 2.0 licence — and that's a big deal. In plain terms, it means anyone can use, modify, and build on GPT-OSS for any purpose, including commercial projects, without paying OpenAI a penny. There are no restrictions on how you distribute your work, and the licence includes patent protections that shield developers from future legal claims.
That makes it one of the most permissive releases from any major AI company. By contrast, Meta's popular Llama models come with brand requirements, and some of Alibaba's Qwen models restrict commercial use for larger organisations.
Carr called it a "maximally permissive licence" and "a very positive, very surprising development."
What this means for smaller players
This is where the story gets genuinely exciting. Until now, running top-tier AI typically meant renting expensive cloud computing or paying per-query API fees. GPT-OSS changes the equation.
A small business, a university research lab, or a solo developer can now run a model rivalling OpenAI's commercial offerings on their own hardware. Data never leaves their machine. There are no ongoing costs beyond electricity. For organisations handling sensitive information — medical records, legal documents, student data — that local deployment option isn't just convenient; it's transformative.
Microsoft moved quickly to make GPT-OSS available through Azure, while AMD and Nvidia both confirmed day-one hardware support. Tools like Ollama and LM Studio offered immediate compatibility, meaning non-technical users could interact with the models through a simple chat interface within hours of launch.
A strategic pivot for OpenAI
The release marks OpenAI's first open language models since GPT-2 in 2019 — back when the company was still a relatively obscure research lab. The decision to return to openness, after years of keeping its most powerful models firmly behind closed doors, signals a shift in competitive strategy.
With China's DeepSeek and other open-source projects gaining ground, and Meta's Llama models establishing a strong foothold, OpenAI may have concluded that keeping everything proprietary was ceding too much territory. GPT-OSS lets the company compete on the open frontier while still selling premium access to its most advanced systems like GPT-5.
Not everyone is fully satisfied. Critics, including researchers at the AI2 institute, note that "open weights" is not the same as truly open source — OpenAI has not released the training data or methods needed to reproduce the models from scratch.
That's a fair point. But for the millions of developers and organisations who simply want to run capable AI locally and affordably, GPT-OSS represents a genuine step forward in making powerful technology accessible to all.
